Cell-Extracellular Matrix Feedback Results in Spontaneous Cell Polarization and Heterogeneous Remodeling in 3D Isotropic and Aligned Discrete-Fiber Models of Cell-Mediated Remodeling Authors: Adam W. Y. Ley, Lauren M. Bersie-Larson, Sabin Adhikari, Paolo P. Provenzano, Kevin D. Dorfman, Victor H. Barocas Published: 2026 ABSTRACT PurposeThe extracellular matrix (ECM) is a dynamic fiber environment containing structural information that significantly impacts cell behavior. Recent experimental work has demonstrated that cells also have significant capacity to remodel their microenvironment, often resulting in ECM heterogeneity. We present an open-source molecular dynamics platform that simulates cell-mediated remodeling wherein cells plasticly remodel their microenvironment and respond to induced structural heterogeneity over multiple retraction cycles.MethodsThe model applies a coarse-grained discrete fiber approach to cell-mediated remodeling. The ECM, represented by a bead-spring polymer, allows proximity-mediated fiber-fiber interactions, representing fiber crosslinking and entanglement. A simulated cell interprets the heterogeneity of its local microenvironment with variable sensitivity and exhibits anisotropic behavior informed by its microenvironment. The cells generate tractors, representing pseudopods, that bind to the ECM and retract towards the cell surface, causing ECM displacement. The cell detaches from the ECM by deleting tractors and allowing the ECM to relax before re-interpreting its surroundings and repeating this process. Metrics of ECM remodeling (fiber densification, orientation, bond strain) and cell morphology were recorded throughout the simulation. The model was extended to a cell remodeling fiber networks with different levels of pre-existing alignment.ResultsThe addition of plasticity in the model enables measurable remodeling: increasing fiber density close to the cell, reorienting fibers radially, and increasing residual fiber bond strain over time. These patterns of remodeling were consistent with previously published experimental results. In initially unaligned fiber networks, cell remodeling resulted in ECM heterogeneity that depended on distance from the cell surface and alignment with the cell’s primary axis. In aligned networks, pre-alignment and sensitivity synergized to increase the heterogeneity of the remodeled networks at further distances from the cell surface.ConclusionsThese findings suggest that cell-ECM feedback mechanisms contribute to heterogeneous remodeling patterns and illustrate that pre-existing alignment impacts remodeling patterns far from the cell. Further, the model presented herein provides a novel modular platform for further investigations into cell-ECM sensing and ECM remodeling heterogeneity.Supplementary InformationThe online version contains supplementary material available at 10.1007/s12195-026-00922-0. Introduction The extracellular matrix (ECM) is a complex microenvironment comprised of fibrous structural proteins like collagen [1]. Early work established that ECM fiber organization induces tissue-scale mechanical anisotropy [2], with higher observed stiffness in the direction of fiber alignment. Aligned structures also have cellular-level impacts, causing cells to preferentially polarize and migrate along pre-existing aligned structures [3, 4]. Such behavior, termed contact guidance [5], is posited to arise from local mechanical anisotropy stemming from structural anisotropy. The ability of cells to reorganize the ECM has also been recognized, particularly in seminal in vitro work from the 1980s [6–8]. Further work has reinforced this view and increasingly recognized the importance of directionally-mediated cellular processes [9, 10]. Pairs of contractile cells generate highly organized, dense structures [11] of aligned fibers [12], which may enable long-distance force transmission between cells [12, 13]. Similarly, the presence of these aligned structures alter cell behaviors, causing preferential migration along aligned fibers and even increasing cell migration speed [14, 15]. As such, there is a need for models that incorporate the development of aligned structures and the effects they have on cells and the microenvironment. Early computational models of ECM remodeling used continuum-based approaches to describe feedback between cells and the surrounding ECM. In the 1990s, Barocas and Tranquillo [16] developed an anisotropic biphasic theory that described cell-mediated ECM remodeling via an ECM-informed model of anisotropic traction force generation. Other continuum models continue to be developed to describe the phenomena in applications including angiogenesis [17, 18] and cell migration [19]. Continuum models cannot, however, account for non-affine deformations [20, 21] characteristic of fibrillar networks and do not allow for analysis of cell-scale microstructural changes. Discrete-fiber computational models of ECM have emerged as a means to understand non-affine fiber network deformations. Many models [22–24] are constrained to two dimensions by computational limitations. Recent three-dimension models have incorporated elements of fiber-fiber interactions [25, 26], such as fiber entanglement and crosslinking. Other models enable interrogation of network topology [27], facilitating quantitative characterization of mechanical and steric ECM anisotropy. There is, however, still need for 3D models that incorporate ECM remodeling and interrogation, in which cells generate and respond to dynamic microenvironmental changes. Further, many existing models make use of in-house codes and are not readily adaptable to parallelization or cluster-based computation. To address this need, we developed a 3D discrete-fiber model of cell-mediated ECM compaction that can integrate fiber-fiber binding and cell-ECM interactions in an open-source format that can be readily customized and integrated into multiscale pipelines using cluster-based computing. Methods Model Development The model is comprised of three elements: (i) a self-interacting bead-spring representation of ECM; (ii) an ellipsoidal cell that generates contractile elements, called “tractors,” representing pseudopodia; and (iii) specified cell response to ECM alignment. Throughout the simulation, tractors were generated in the ECM far from the cell, bound to fibers, retracted towards the cell via controlled displacement, and finally detached from ECM. Fibers that contacted during the simulation formed new bonds, resulting in plastic changes to ECM topology. Network Generation ECM networks were generated by constructing a Voronoi or Delaunay tessellation on uniformly random seeded points in a computational volume using Python’s built-in Voronoi and Delaunay tessellation codes. The resulting tessellation edges were then treated as fibers and the intersections as pre-existing crosslinks within the fiber network. The fiber volume fraction φ (Eq. 1) of the network was then determined by1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\upvarphi =\frac{\sum_{f}{l}_{f}\pi {r}^{2}}{V}$$\end{document}φ=∑flfπr2Vwhere f are fibers in the network, lf is the length of each fiber, V is the volume of the simulated space, and r is the radius of the simulated fibers, taken to be 100nm [28]. All networks were generated to have a volume fraction similar to a 1 mg/mL collagen hydrogel [25]. The generation volume typically had 8x the volume of the intended simulation volume to mitigate edge effects. Alignment in networks can be induced by stretching the network in the coordinate directions. Stretching the network in one direction produces linearly aligned networks while stretching in two directions can create planarly aligned networks. In this study, uniaxial stretch ratios of 1, 1.05, 1.1, 1.25, 1.5, and 2 were explored. ECM parameters and plasticity Tessellation edges were modeled as cylindrical ECM fibers of 100 nm radius [28], and tessellation nodes formed pre-existing fiber crosslinks. Networks were cropped into 100-μm spheres that placed initially spherical cells 90 μm, or 4.5 cell-lengths, away from the simulation boundaries. Aligned networks were generated similarly but were aligned by stretching in the x-direction by the their stretch ratio before cropping. Delaunay-tessellation-based networks were also generated to explore how topological differences alter ECM remodeling. Each tessellation edge was then segmented into equally spaced segments by dividing the edge length by 400 nm and rounding to the next integer. For instance, a 1700 nm edge would be divided into five equal 340 nm segments. Beads representing ECM fibers were placed along the tessellation edge and connected by harmonic springs (Fig. 1a) (Eq. 2),2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E}_{bond}(r)=K{\left(r-{r}_{0}\right)}^{2}$$\end{document}Ebond(r)=Kr-r02where Ebond is the bond energy, K is the spring constant, r is the distance between beads, and r0 is the bond rest length, 400 nm. This method put the fibers in the network under initial compression that was later relaxed to form a network that had no initial bond strain. Since the tractors move via controlled displacement, the units of K are arbitrary in nature if the focus is on kinematics and not force generation. Bending of fiber segments was not included in this simulation set; when the fibers were loaded in compression, fiber segments were able to buckle up to the point that two beads came within a fiber diameter (200 nm) of one another but then repulsed according to pair-wise interaction detailed next.Fig. 1a Network edges were segmented into equidistant beads and connected via harmonic springs. b Beads from different fibers are prevented from overlapping via Weeks-Chandler-Anderson potentials. c In plastic simulations, contacting fibers form new crosslinks. d Schematic of cell-driven remodeling. a Network edges were segmented into equidistant beads and connected via harmonic springs. b Beads from different fibers are prevented from overlapping via Weeks-Chandler-Anderson potentials. c In plastic simulations, contacting fibers form new crosslinks. d Schematic of cell-driven remodeling. All ECM beads interacted pairwise via a Weeks-Chandler-Anderson (WCA) potential [29] (Fig. 1b) (Eq. 3),3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E}_{pair}(r)=\left\{\begin{array}{ll}4\varepsilon \left[{\left(\frac{\sigma }{r}\right)}^{12}-{\left(\frac{\sigma }{r}\right)}^{6}\right], &\quad r<{r}_{cf}\\ 0, &\quad r\ge {r}_{cf}\end{array}\right.$$\end{document}Epair(r)=4εσr12-σr6,r 0 increased in semimajor axis length (Fig. 10, first column), but the polarization direction was random. Thus, sensitivity increased the magnitude of cell polarization in initially unaligned gels, but it did not predispose cells to orient in a particular direction. When pre-existing alignment was introduced into the networks, cells of non-zero sensitivity aligned in the x-direction. Cell sensitivity increased both the degree of polarization and directional preference.Fig. 10The direction of each cell’s major axis in the x-y plane plotted vs the cell’s longest axis after 40 retractions (n = 10). Each arrow represents a different cell among 10 different runs. Sensitivities: (a-d) s = 1, (e-h), s = 2, (i-l) s = 6. Initial ratio of fiber alignment in x, y, z coordinate directions: (a,e,i) 1:1:1. (b,f,j) 1.08:1:1, (c,g,k) 1.5:1:1, (d,h,l) 2:1:1. Dispersity decreased with both sensitivity and initial alignment. The direction of each cell’s major axis in the x-y plane plotted vs the cell’s longest axis after 40 retractions (n = 10). Each arrow represents a different cell among 10 different runs. Sensitivities: (a-d) s = 1, (e-h), s = 2, (i-l) s = 6. Initial ratio of fiber alignment in x, y, z coordinate directions: (a,e,i) 1:1:1. (b,f,j) 1.08:1:1, (c,g,k) 1.5:1:1, (d,h,l) 2:1:1. Dispersity decreased with both sensitivity and initial alignment. Effect of pre-alignment Aggregate ECM densification was quantified at different distances for different combinations of cell sensitivity and network pre-alignment. Near the cell (Fig. 11a), densification decreased with increased pre-alignment and cell sensitivity. Intermediate distances (Fig. 11b) exhibited a similar densification pattern with reduced magnitude, but farther distances (Fig. 11c) did not.Fig. 11a–c Average fiber volume fractions (n = 10) after 40 retractions plotted against the initial network x-alignment and cell sensitivity. a 7.5 μm, b 12.5, and c 27.5 μm from the cell surface. d–f Fitted von Mises κ parameters for fiber density plotted against the initial network x-alignment and cell sensitivity (d) 7.5 μm, e 12.5, and f 27.5 μm from the cell. a–c Average fiber volume fractions (n = 10) after 40 retractions plotted against the initial network x-alignment and cell sensitivity. a 7.5 μm, b 12.5, and c 27.5 μm from the cell surface. d–f Fitted von Mises κ parameters for fiber density plotted against the initial network x-alignment and cell sensitivity (d) 7.5 μm, e 12.5, and f 27.5 μm from the cell. Directional data at each distance were fitted with von Mises distributions. Near the cell (Fig. 11d), the κ parameter increased with both sensitivity and network pre-alignment. Farther from the cell (Fig. 11e,f), the same correlation emerged but with lower magnitude. Thus the heterogeneity signal remains Ω- and s-dependent at further distances while the aggregate effects decay. Discussion By incorporating elements of cell-ECM sensing and ECM plasticity, the model presented herein recapitulates certain experimentally-observed trends in cell-mediated ECM remodeling. The model exhibits a distance-dependent densification pattern that decreases in rate with subsequent retraction cycles, qualitatively matching published experimental results [11]. The cell model also captures long-observed trends of cell polarization behavior: persistent polarization [34] and strong polarization responses in anisotropic 3D ECMs [4, 30]. Qualitative observations of remodeled ECM indicate that fibers near the cell reorient either radially about the cell center or linearly with respect to the cell’s polarization direction [4, 35]. The model reproduces this trend, with less sensitive cells exhibiting greater radial fiber reorientation and more sensitive cells exhibiting greater linear fiber reorientation. These combined trends of increasing density and linear alignment are often associated with network stiffening or mechanical anisotropy [36]. Thus, the ability to examine both trends separately may provide further insight into microstructural mechanical anisotropy. Further, the model enables investigation of the effect of existing ECM alignment on remodeling and ECM strain. Previous simulations in discrete networks have demonstrated that a small proportion of fibers are loaded during cell retraction [24, 25] and that these fibers can transmit forces across long distances [25]. The model recapitulates this trend and demonstrates that directional strain fields arise in both initially isotropic ECM networks remodeled by high-sensitivity cells or less sensitive cells in aligned networks. Experimentalists have developed numerous ways of creating aligned fiber networks for applications from simulating the tumor microenvironment [37, 38] to developing regenerative medicine platforms [39]. Such advances make it essential that models incorporate a cell-ECM sensing mechanism to investigate how remodeling patterns change with either pre-existing or evolving anisotropy. The current modeling approach includes simplifications that may warrant future investigation. The boundary conditions of the model may make the results sensitive to edge effects. In experiments, collagen gels are observed to undergo large amounts of compaction [40]. Currently, the model maintains a constant volume with fixed boundaries, and as a result it may not be suitable for simulating long-term remodeling behavior or highly cellularized tissues. This simplification in the boundary conditions may also prevent good quantitative matches for both the ECM network densification and realignment, as fibers nearer to the boundaries are unable to displace towards the cell or propagate stresses beyond the simulation volume. Additionally, previous simulations in discrete fiber networks have demonstrated that a small proportion of fibers transmit significantly higher forces across long distances [24, 25]. Especially in fibrous environments, non-affine deformations are often necessary for recapitulating the displacement and behavior of this small but relevant population of the network [20]. While both discrete and continuous 2D models can recapitulate force transmission anisotropy [22, 24], existing models have difficulty recapitulating the formation of long-range stress cords due to their continuum nature or constraint in the z-axis. The 3D discrete fiber model captures observed long-range force transmission anisotropy [13] and can provide insight into how aligned structures arise in initially unaligned fiber environments, however these long-range fiber strains quickly propagate to the simulation boundaries in relatively few retractions in high sensitivity cells and cells in highly aligned environments. As such, the model’s ability to measure the length-scale of these signals may be limited by the model’s boundary conditions. Further, these results suggest that the model is likely sensitive to a number of network parameters influencing force propagation, such as the average connectivity of nodes within the ECM as well as the ratio between the initial fiber lengths and both the retraction distance and the length-scale of the simulation volume, which were not quantified in this study. The model’s use of prescribed tractor displacement is likely sufficient for small-scale deformations in the material as each total retraction displaces the ECM 5% the length scale of the volume over the course of 50 steps. However, displacement-controlled tractors may not fully approximate the behavior of stress-controlled tractors nor reflect changes in changes in cell contractility. Experimental data suggest that the contractility behavior [11, 41] are highly dependent ECM density and linear alignment, which both affect ECM stiffness perceived by the cell [36]. While the model simulates changes in cell contractility by increasing tractor density at the polar extents of the cell, contractility could also be modulated by changing the size of the tractors as well as their retraction distance. As such, measurements of ECM remodeling may be sensitive to both the size of tractors as well as their contractility, however, this work did not quantify such sensitivities. The model also interrogates ECM structural anisotropy instead of a perceived stiffness around the cell. While previous work has demonstrated cells exhibit contact guidance behavior even with disrupted integrin binding [30], the model does not currently account for perceived stiffness anisotropy. This perceived stiffness is likely different than bulk measurements measured via macroscale experiments, however, smaller micro-and nano-scale investigations are starting to lend insight into cell-scale ECM perception [42–44]. ECM stiffness may also impact how cells establish tensional homeostasis, which is implicated in remodeling processes [10, 25, 45–47]. As such, a force-dependent tractor model may be able to capture force-dependent phenomena more accurately. Further, the sensitivity expression and the tractor extension and retraction mechanisms are arbitrary and do not account for the considerable literature on the inner workings of the cell, especially contractility [48, 49]. While computational tractability could become an issue, the model could be further developed to incorporate cytomechanical models, like the motor clutch model [50, 51], or additional ECM interrogations, such as local stiffness and steric hinderance [27, 52]. The current modeling framework also lacks an inherent timescale largely due to its approach in tractor retraction. Though it is possible to measure the rates of pseudopod extension and retraction [53], the model simplifies this action by conducting a single large contraction, effectively summing the action of multiple extension and retraction events. Thus the 40 tractors modeled in each retraction event is likely larger than the average number of pseudopod extensions maintained by a cell. These parameters were selected to elicit remodeling behaviors in a computationally tractable manner. The relaxation of the network during after retraction and detachment is also non-physiological and ignores sustained cellular tension. While the current approach can approximates ECM densification patterns and anisotropic strain propagation, time-resolved approaches where tractors probabilistically bind and unbind the ECM may lend additional insight dynamic tension generation, especially if paired with cytomechanical models for actin tension generation [50, 51]. The lack of an apparent timescale may also influence the ECM’s crosslinking behavior. While fiber reorientation and displacement resolve as fast as the retraction process, crosslinking and entanglement may occur on a longer timescale. Further approaches that break up the mass retraction process into individual time-resolved events and a probabilistic model for fiber-fiber bonding may be able to account for these in future models. Finally, this work presents a customizable framework that can increase its functionality by leveraging existing software packages and LAMMPS’ built-in functions. While the presented work used Voronoi and Delaunay tessellations, additional network topographies could be simulated [25]. Further, while the networks were simulated as networks of homogenous fibers, fiber-specific parameters such as diameter and stiffness could be altered upon network generation by defining different molecule types and different rules of interaction. These interaction rules could be further extended to prevent fibers of a certain type from binding to other fibers or to change universal crosslinking rules. Additional fiber properties like bending stiffness can be added to fiber segment definitions, conferring additional resistance to bending within simulated networks. Further, fiber segments can be repeatedly updated between retraction steps to change to fiber spring constants, resting distances, and Lennard-Jones length constants to simulate the fiber stiffening or thickening dependent on the stress state of fibers or fiber segments. However, the creation of many different types of fibers and additional rules for interaction would necessarily add to the computational complexity and memory demands of such simulations. Finally, the rules governing fiber-fiber interactions can be further extended. While fiber-fiber crosslinks formed deterministically in the presented results, fiber crosslinking can also be probabilistically determined, promoting bond creation between fibers that contact each other longer than transient interactions. Bond breakage can also be handled by existing functions in LAMMPS, enabling stress-dependent bond breakage within the fiber network. Bond breakage was not implemented in these experiments due to the nature of the static cell and the potential detachment of the network nearest the cell from the rest of the ECM. The addition of such parameters would increase the computational cost of simulations, but LAMMPS’ parallelizable structures and GPU implementations may be leveraged to increase computational tractability. Despite these shortcomings, the model presents an open-source and expandable framework that synthesizes elements of cell ECM-feedback mechanisms and ECM plasticity to provide insight into ECM heterogeneity at the fiber level. The current implementation of the model can capture experimentally observed cell and ECM remodeling trends and the model’s modular nature will enable further development in this framework with adjustable methods for cell ECM sensing, cell sensitivity, and tractor manipulation. Finally, the model’s implementation in an open-source molecular dynamics framework lends itself to parallelization and cluster-based computing. Electronic supplementary material Below is the link to the electronic supplementary material.Supplementary file1 (DOCX 2764 kb) Supplementary file1 (DOCX 2764 kb)